OFDM channel estimation method based on mamba neural network
By utilizing the parallel computing and hidden state compression techniques of the Mamba neural network architecture, the problems of low signal-to-noise ratio and high complexity in traditional channel estimation methods are solved, achieving accurate estimation and fast response of channel state, which is applicable to 4G and 5G mobile communication systems.
Patent Information
- Application Number
- CN202511225641.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-29
- Publication Date
- 2025-11-14
AI Technical Summary
Existing channel estimation methods are susceptible to interference under low signal-to-noise ratio conditions, have high computational complexity, and traditional deep learning models such as RNNs and attention mechanisms suffer from low training efficiency and difficulty in modeling out-of-window sequences.
By employing the Mamba neural network architecture and utilizing parallel computing and hidden state compression techniques, the window limitation is overcome to achieve accurate estimation of long-term CSI sequences. Furthermore, error observation and network parameter updates are performed by inserting CSI at pilot positions to optimize estimation accuracy.
It achieves efficient and low-complexity CSI estimation in rapidly changing channel environments, accurately tracks channel state changes, improves training efficiency, and reduces computational complexity.
Smart Images
Figure CN120956569A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an OFDM channel estimation method based on a Mamba neural network, belonging to the technical field of channel estimation methods. Background Technology
[0002] Orthogonal Frequency Division Multiplexing (OFDM) technology is a key technology in 4G and 5G mobile communication systems. In wireless communication systems, transmitted signals are affected by the characteristics of the propagation channel, resulting in distortion, fading, and path loss. Received signals are also damaged by additive noise and interference, further degrading signal quality. Accurate CSI estimation is fundamental for effective signal processing and data recovery.
[0003] Traditional channel estimation methods primarily rely on pilot information. This approach involves two steps: First, fixed pilot symbols, pre-agreed upon with the receiver, are inserted at the transmitter in a specific order. The receiver extracts the symbols at the pilot positions from the received signal and compares them with the original transmitted pilot symbols to obtain the Channel State Indicator (CSI) at those positions. Then, based on the CSI at the pilot positions, an interpolation algorithm is used to obtain the CSI at the data positions, ultimately yielding the CSI at all symbol positions. The structure of the inserted pilots is mainly divided into three categories: block pilots, comb pilots, and lattice pilots. The most common interpolation algorithms include Least Squares (LS) and Minimum Mean Square Error (MMSE). LS is simple to implement, but it doesn't consider noise during estimation, making it susceptible to interference and noise at low signal-to-noise ratios (SNR), limiting its applicability to a narrow range of SNRs. MMSE, on the other hand, utilizes second-order channel statistics and noise, resulting in better channel state estimation. However, it is more complex, requiring significant computational resources, and cannot be used when channel statistics are unavailable.
[0004] Deep learning technology has attracted attention from scholars in various fields due to its powerful feature extraction and pattern fitting capabilities. In using deep learning technology to solve the channel state estimation problem, the main solutions fall into two categories.
[0005] One type of approach analogizes the channel estimation problem to the super-resolution upsampling problem in image processing. These methods primarily utilize convolutional neural networks (CNNs) to construct image super-resolution and denoising networks. In the first stage, the channel estimation matrix, containing only pilot position channel state information, is treated as an image, and upsampling is performed using the super-resolution network. In the second stage, the denoising network filters out noise. This CNN-based approach has shown some success in channel estimation, but it doesn't utilize the long-term correlations of temporal frequency variations in the channel matrix. The second type of approach builds upon CNNs by further utilizing recurrent neural networks (RNNs) to capture temporal correlations. Systems based on long short-term memory (LSTM) recurrent neural networks have proven superior to purely CNN-based systems. Methods based on gated recurrent units (GRUs), using computationally less complex RNN blocks, have also proven superior to LSTM-based systems. Attention mechanisms are effective architectures validated by many large models, and their application in CSI estimation is also a key research focus. However, RNNs require knowledge of the hidden state from the previous time step to generate output. Therefore, during training, RNNs cannot perform parallel computation and must proceed sequentially, resulting in low training efficiency. Attention-based neural networks can achieve parallel computation during training, but they are still limited by window length and cannot model sequences outside the window. Furthermore, during channel state estimation, each estimation requires recalculating all hidden states, increasing computational complexity quadratically with the window length. Therefore, achieving efficient training and low-complexity CSI estimation for deep learning-based CSI estimation models has become a new challenge. Summary of the Invention
[0006] The purpose of this invention is to disclose an OFDM channel estimation method based on a Mamba neural network. This method overcomes the window limitation of attention mechanisms and efficiently utilizes the internal correlation of long-term CSI sequences to achieve accurate channel state estimation. During training, parallel computing can be employed to improve training efficiency. When estimating CSI, only the current input and the hidden state from the previous time step need to be considered, effectively reducing the computational complexity of estimation and achieving fast estimation. This method can update the neural network parameters as needed during CSI estimation, optimizing the model's estimation accuracy in a timely manner. This allows for more effective handling of rapidly changing channel environments with dual selective fading.
[0007] To achieve the objectives of this invention, this invention provides an OFDM channel estimation method based on a Mamba neural network, which includes the following steps: Step 1: Build a neural network using the Mamba architecture and train it until the loss function converges; Step 2: Obtain the channel state matrix of the Nth time slot containing only pilot positions by inserting pilot signals. , and The inputs are fed column by column into the trained neural network, resulting in matrices of the same dimension formed by the network's outputs. ; Step 3: Using matrices Data from the mid-frequency position is used to replace the output matrix of the neural network. By analyzing the data at the corresponding locations, a complete channel state estimation matrix can be obtained. ; Step 4: Using functions calculate and If the difference between the pilot frequency position data and the value exceeds a set threshold, the neural network parameters will be updated. Step 5: Based on the channel state matrix of the pilot position in the subsequent time slots, repeat steps 2 to 4 to obtain the channel state estimation matrix for each subsequent time slot.
[0008] Preferably, in step one, the data processing flow of the constructed Mamba architecture neural network is as follows: the input is a channel state matrix containing only pilot positions. , like Figure 1 The position of the black dot shown indicates that its row number is The number of columns is , indicating the first line Time within a time slot The columns represent different communication frequencies. First, the data is processed column by column through a convolutional neural network. Transform into columns of the same dimension ;Then After passing through a fully connected neural network, P1, P2, and P3 are mapped to matrices B, C, and discretized parameters, respectively. ; Utilizing discretized parameters According to the discretization rule , Discretize matrices B and A to obtain and .
[0009] Preferably, the discretization rule is the zero-order preservation rule:
[0010] According to the formula , , will get Input to a fully connected neural network Get output ,use replace The channel estimation at the middle pilot position is obtained as follows: Channel estimation of the entire communication sub-channel Other sub-channels can be obtained simultaneously using a parallel approach. Channel estimation is performed to obtain the current time slot. Channel characteristic estimation .
[0011] Preferably, the matrix The construction uses the Hippo matrix construction rules. .
[0012] Preferably, in step one, the neural network training process of the Mamba architecture includes the following steps: 1) Constructing the training dataset , ,… , ,in, To obtain a matrix containing only the channel states at the pilot positions by inserting pilots, This is the actual channel state matrix; 2) Randomly select from the training data A continuous matrix The input is fed into the neural network to obtain the output. ; 3) Utilizing the loss function ,in The difference between the two is represented by the loss function, which updates the parameters of the neural network, including convolutional neural networks and fully connected networks P1, P2, P3 and Q. 4) Repeat steps 2 and 3 until the loss function is calculated. convergence.
[0013] Preferably, the loss function is ,in .
[0014] Preferably, the function Same as the loss function.
[0015] Preferably, the training dataset , ,… , middle, .
[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention can overcome the window limitation of the attention mechanism and efficiently utilize the internal correlation of long-term CSI sequences to achieve accurate estimation of channel state.
[0017] Neural networks employing attention mechanisms have a window limit on the length of the CSI sequence, preventing the modeling of data outside the window and thus hindering the estimation of the current channel state. The proposed solution eliminates this window limit on the length of the data used for estimation and utilizes hidden states to compress historical data. Therefore, the estimation of the current CSI is based on all previous historical data, rather than the data sequence within the window limit, resulting in a more accurate estimation of the CSI.
[0018] 2. In the training process, the present invention can improve training efficiency by using parallel computing.
[0019] This scheme uses hidden states to compress historical CSI sequence data, thereby breaking through the window limitation of data length. However, compared with RNN neural networks that also use hidden states, the proposed scheme does not need to know the previous state to calculate the next state during training. Instead, it can achieve parallel computation through the parallel scanning algorithm in Mamba, thus improving training efficiency.
[0020] 3. When estimating CSI, this invention effectively reduces the computational complexity of the estimation and achieves rapid estimation.
[0021] After training, this scheme, when estimating CSI, does not require calculating the correlation of all data within the window, unlike neural networks based on attention mechanisms. The proposed scheme only needs to consider the current input and the hidden state from the previous time step to estimate the current channel state data, resulting in low complexity and high inference efficiency.
[0022] 4. This invention can update the parameters of the neural network as needed during the CSI estimation process, optimize the estimation accuracy of the model in a timely manner, and track changes in channel state.
[0023] Since the CSI at the inserted pilot position is available in practical applications, this scheme can also be used to compare the CSI at the inserted pilot position with the CSI estimated by the network. The difference between the two can be used to observe the estimation error and update the network parameters. The combination of basic training and network fine-tuning can be used to track changes in channel state and optimize the estimation accuracy. Attached Figure Description
[0024] Figure 1 This is a flowchart of the data processing of the Mamba neural network of the present invention. Detailed Implementation
[0025] The present invention will be further described in detail below with reference to embodiments, so that those skilled in the art can implement it based on the description.
[0026] Orthogonal Frequency Division Multiplexing (OFDM) is a key technology in 4G and 5G mobile communication systems. Fading characteristics in wireless channels often degrade communication signal quality and affect communication performance. Accurate signal identification and signal recovery (CSI) estimation is fundamental for effective signal processing and data recovery. Deep learning-based CSI estimation algorithms, as alternatives to traditional algorithms, have achieved significant research results. However, overcoming the limitations of RNNs (low training efficiency due to inability to perform parallel computation), the inability of attention mechanisms to model sequences outside the window, and the fact that estimation complexity increases quadratically with the window length, remains a major challenge in utilizing deep learning for CSI estimation.
[0027] In order to solve the problems existing in the prior art, in the embodiments disclosed in this invention: First, construct a neural network using the Mamba architecture and train it accordingly, referring to the appendix. Figure 1 In the constructed Mamba neural network, the data processing flow is as follows: the input is a channel state matrix containing only pilot positions. , Figure 1 The position of the black dot shown indicates that its row number is The number of columns is , indicating the first line Time within a time slot The columns represent different communication frequencies. After processing by a convolutional neural network Transform into columns of the same dimension ;Then After passing through a fully connected neural network, P1, P2, and P3 are mapped to matrices B, C, and discretized parameters, respectively. ; Utilizing discretized parameters According to the discretization rule , Discretize matrices B and A to obtain and .
[0028] Preferably, the discretization rule is the zero-order preservation rule.
[0029] According to the formula , , will get Input to a fully connected neural network Get output ,use replace The channel estimation at the middle pilot position is obtained as follows: Channel estimation of the entire communication sub-channel Other sub-channels can be obtained simultaneously using a parallel approach. Channel estimation is performed to obtain the current time slot. Channel characteristic estimation .
[0030] matrix The construction uses the Hippo matrix construction rules. .
[0031] Second, by inserting pilots, the channel state matrix containing only pilot positions in the Nth time slot is obtained, and then the matrix is input column by column into the trained neural network to obtain the corresponding output. Third, utilize In Replace the output of the neural network Channel estimation data at the middle pilot position is used to obtain the channel state estimation matrix. ; Fourth, using functions calculate and If the difference in pilot position data exceeds a set threshold, the neural network parameters will be updated. Fifth: Based on the channel state matrix containing only pilot positions in the subsequent time series, repeat steps two to four to obtain the channel state estimation matrix for each subsequent time slot in turn.
[0032] This invention requires training a Mamba neural network, specifically including the following steps: 1) Constructing the training dataset , ,… , ,in, To obtain a matrix containing only the channel states at the pilot positions by inserting pilots, This is the actual channel state matrix; 2) Randomly select from the training data A continuous matrix The input is fed into the neural network to obtain the output. ; 3) Utilizing the loss function ,in The difference between the two is represented by the loss function, which updates the parameters of the neural network, including convolutional neural networks and fully connected networks P1, P2, P3 and Q. 4) Repeat steps 2 and 3 until the loss function is calculated. convergence.
[0033] The loss function is: ,in .
[0034] function Same as the loss function. Training dataset. , ,… , middle, .
[0035] This embodiment overcomes the window limitation of attention mechanisms, efficiently utilizing the internal correlation of long-term CSI sequences to achieve accurate channel state estimation. During training, parallel computing can be employed to improve training efficiency. When estimating CSI, only the current input and the hidden state from the previous time step need to be considered, effectively reducing computational complexity and enabling rapid estimation. This method can update neural network parameters as needed during CSI estimation, optimizing the model's estimation accuracy in a timely manner. This allows for more effective handling of rapidly changing channel environments with dual selective fading.
Claims
1. An OFDM channel estimation method based on a Mamba neural network, characterized in that... Includes the following steps: Step 1: Build a neural network using the Mamba architecture and train it until the loss function converges; Step 2: Obtain the channel state matrix of the Nth time slot containing only pilot positions by inserting pilot signals. , and The inputs are fed column by column into the trained neural network, resulting in matrices of the same dimension formed by the network's outputs. ; Step 3: Using matrices Data from the mid-frequency position is used to replace the output matrix of the neural network. By analyzing the data at the corresponding locations, a complete channel state estimation matrix can be obtained. ; Step 4: Using functions calculate and If the difference between the pilot frequency position data and the value exceeds a set threshold, the neural network parameters will be updated. Step 5: Based on the channel state matrix of the pilot position in the subsequent time slots, repeat steps 2 to 4 to obtain the channel state estimation matrix for each subsequent time slot.
2. The OFDM channel estimation method based on a Mamba neural network according to claim 1, characterized in that, In step one, the data processing flow of the constructed Mamba architecture neural network is as follows: the input is a channel state matrix containing only pilot positions. As shown in Figure 1, the black dot position has a row number of 1. The number of columns is , indicating the first line Time within a time slot The columns represent different communication frequencies. First, the data is processed column by column through a convolutional neural network. Transform into columns of the same dimension ;Then After passing through a fully connected neural network, P1, P2, and P3 are mapped to matrices B, C, and discretized parameters, respectively. ; Utilizing discretized parameters According to the discretization rule , Discretize matrices B and A to obtain and .
3. The OFDM channel estimation method based on a Mamba neural network according to claim 2, characterized in that, The discretization rule is a zero-order preservation rule: ； According to the formula , , will get Input to a fully connected neural network Get output ,use replace The channel estimation at the middle pilot position is obtained as follows: Channel estimation of the entire communication sub-channel Other sub-channels can be obtained simultaneously using a parallel approach. Channel estimation is performed to obtain the current time slot. Channel characteristic estimation .
4. The OFDM channel estimation method based on a Mamba neural network according to claim 2, characterized in that, The matrix The construction uses the Hippo matrix construction rules. .
5. The OFDM channel estimation method based on a Mamba neural network according to claim 2, characterized in that, In step one, the neural network training process of the Mamba architecture includes the following steps: 1) Constructing the training dataset , ,… , ,in, To obtain a matrix containing only the channel states at the pilot positions by inserting pilots, This is the actual channel state matrix; 2) Randomly select from the training data A continuous matrix The input is fed into the neural network to obtain the output. ; 3) Utilizing the loss function ,in The difference between the two is represented by the loss function, which updates the parameters of the neural network, including convolutional neural networks and fully connected networks P1, P2, P3 and Q. 4) Repeat steps 2) and 3) until the loss function is calculated. convergence.
6. The OFDM channel estimation method based on a Mamba neural network according to claim 5, characterized in that, The loss function is: ,in .
7. The OFDM channel estimation method based on a Mamba neural network according to claim 5, characterized in that, The function Same as the loss function.
8. The OFDM channel estimation method based on a Mamba neural network according to claim 5, characterized in that, The training dataset , ,… , middle, .